😸 Google's secret weapon

Welcome, humans.

It’s Friday, it’s been a long week, and we know you’re tired of working. So in order to celebrate the weekend, and get things rolling, here is a website that rates how well your cat sits like loaves of bread:

Move over Simone Biles, because Nougat just got a 99 out of 10. Forget the GOAT, it’s all about the LOAF.

Just upload a picture of your cat sitting like a loaf, and you get a score. Simple as that.

Don’t you just loaf the internet?

ICYMI: We’re partnering with Exec on a ChatGPT Bootcamp that’ll 2x your team's productivity. Limited spots are available, so book a consultation here.

Here’s what you need to know about AI today:

• Google’s experimental new Gemini model topped the main AI leaderboard.
• OpenAI announced a GPT-4o “long context” version that has a 64,000-token output.
• Google added new AI features to Chrome, including improved Lens and tab comparison.
• Suno defended against record labels' lawsuit, claiming "learning is not infringement."

Advertise in The Neuron here.

Google's new Gemini 1.5 Pro model is available to test and and topped this AI leaderboard.

The past two weeks have been a whirlwind of AI releases: Meta's new open source models, ChatGPT 4-o mini, Apple Intelligence... Now Google is joining the fray.

Did we save the best for last? Because Google’s new model, Gemini 1.5 Pro "experimental 0801", is currently topping the LMSYS leaderboards as the best chatbot out there.

Ever since the announcement, which surprised most of us, people on X have been throwing their toughest questions at it. It’s done well, but early results seem inconsistent. It is an experiment, after all.

Why's this a big deal? Seven major models have been released in the last three months. Every major player has a frontier model of some kind, and each one is competitive in its own way.

Last week, it was Meta Llama 3.1 topping charts (and popping bottles). This week, Google basically said "hold my beer" and shot to the top with barely a peep. No blog post announcement, just some tweets from the team.

Now, it seems niching down into certain specialty skills (coding, math, multilingual performance, etc) is where the bulk of the performance gains will come from over the next few releases.

Gemini’s niche? Definitely math (wonder why…). It’s ranked #1 at solving math problems, ranked in the top 1-2 at following instructions, and in the top 5 at handling tricky English tasks and coding.

On that note, we couldn't resist throwing it the infamous “Turbo the snail” problem that stumped AlphaProof at the IMO. Gemini took a crack at it, but ehh… we're not exactly math whizzes ourselves. Anyone out there want to fact-check an AI's homework?

What’s our take? We tried it, and it’s definitely an experiment. At times, it's the smartest system we’ve tested; other times, it mistakes a hexagon for a heptagon.

Grant’s chat with the new Gemini model.

But then again, whomst among us hasn’t?

Anyone can test it out for free in Google's AI Studio by selecting model > Gemini 1.5 Pro Experimental 0801, but be warned: this version is for developer feedback specifically.

For now, we’d say 1.5 Pro is perfect for tinkering and testing, but maybe hold off on using it for anything mission-critical—unless you enjoy living dangerously!

FROM OUR PARTNERS

AI as it Should be

One Platform.

Tailored to you. Unlock the real value of AI with local, ownable, and customizable models that you control.


webAI allows you to:


• Test and iterate solutions directly, with no need for a dedicated AI team.
• Utilize local compute, meaning faster, secure testing at lower costs.
• Deploy effortlessly across local, cloud, edge, or hybrid environments.


Learn More

Around the Horn.

• OpenAI will release a new model called GPT-4o Long Output with a 64,000 output window—roughly equivalent to a 200 page novel in length.
• Google launched 3 new AI features for Chrome—improved Google Lens, Tab Compare for reviewing products on multiple tabs, and conversational search for your history.
• The EU’s AI Act went into effect yesterday (August 1st), and U.S. big tech firms could be fined up to $41M (or 7% of global revenues) if they break it.
• Suno accused the record labels suing it of being afraid of competition, admitted to using some copyrighted work in its training, but claimed “learning is not infringement.”
• Argentina will use AI to predict future crimes in a plot eerily familiar to the sci-fi story Minority Report.

The fastest way to build AI apps

• Writer Framework: build Python apps with drag-and-drop UI
• API and SDKs to integrate into your codebase
• Intuitive no-code tools for business users

Start building with AI Studio

Treats To Try.

1. Stable Fast 3D generates 3D assets from a single image (in just 0.5 seconds) for game developers, VR creators, and professionals in design-intensive fields.
2. Runway ML trained a new version of Gen-3 Alpha called Turbo that generates videos 7x faster than the original (sometimes ~11 secs to make a 10 sec video).
3. Black Forest Labs released Flux AI, an open-source text-to-image model to match Midjourney quality, with 3 versions available and plans for future text-to-video models (raised $31M).
4. Clarity records, analyzes, and extracts insights from customer calls to improve sales and product development.
5. EduWiz.AI helps generate essays, research papers, and other content with AI autocomplete, paraphrasing, and customizable document formats.
6. Powder does precise document analysis, automating tedious tasks like proposal building and estate analysis to reduce processing time by 95% (raised $5M).

*This is sponsored content. Advertise in The Neuron here.

Intelligent Insights

• NVIDIA's Jensen Huang and Meta's Mark Zuckerberg discuss AI research, generative AI's impact on developers, and virtual worlds' role in AI advancement (link).
• Learn how AI is becoming a major player at the Olympics (link) and gymnastics specifically (link).
• AI researcher Ethan Mollick compares Advanced Mode to Siri—and shares why he thinks voice mode might be the future UI of AI (link).
• MIT researchers developed a new system that uses a smartphone scan of your home to train robots digitally so they don’t have to break as much stuff learning IRL (link).
• This 76-page report analyzed over 1,500 AI papers to create a roadmap of prompting techniques for AI models, revealing which strategies work best and why. (link)

A Cat's Commentary.